Base Model Training with Knowledge Graph Information Matrices
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models face challenges in object detection, trajectory prediction, and motion planning for vehicles when there is limited or no training data available, particularly in scenarios where objects like road signs are not present in the training data.
Innovation Solution
A method and system for training a base model using a knowledge graph that provides domain-specific knowledge, allowing the model to understand and predict driving scenes by generating information matrices from image data and knowledge graphs, and then training the model on these matrices.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If base models are trained using only available training data without domain-specific knowledge, then the model can be trained with simple data processing, but the model cannot detect objects or scenarios not present in the training data
Solution Approach 1:
The patent applies preliminary action by pre-processing training data into structured information matrices that encode domain-specific relationships and attributes before model training. This preprocessing step creates enriched training representations that enable the model to generalize to new categories without requiring complex architectural changes or extensive retraining
Solution Approach 2:
The patent introduces information matrices as an intermediary between raw training data and the base model. These matrices serve as a structured representation that captures domain-specific knowledge about objects, attributes, and relationships, allowing the model to learn from organized information rather than raw pixels, thereby improving adaptability to new categories
2Adaptability or versatility
If classification approaches use manually created attribute structures to compensate for missing data, then the model can handle objects not in training data, but the manually created structures are inaccurate and difficult to adapt to new areas
Solution Approach 1:
The patent applies dynamics by creating information matrices that can dynamically adapt to different domains and use cases. The structured representation allows attributes and relationships to be flexibly defined and modified based on the specific application area, enabling accurate adaptation to new domains without relying on rigid manual attribute structures
Solution Approach 2:
The patent changes the parameter representation by transforming raw image data into structured information matrices with specific parameters for objects, attributes, and relationships. This parameter transformation enables more precise attribute identification compared to manual approaches, while maintaining adaptability through the structured format that can be customized for different domains
3Reliability
If base models are trained with descriptive captions from web data, then the model can learn from text and images, but the approach cannot be adapted to new categories and does not integrate human domain knowledge
Solution Approach 1:
The patent applies preliminary action by pre-structuring training data into information matrices that encode domain-specific knowledge before model training. This preprocessing creates enriched representations that enable the model to achieve high predictive accuracy while maintaining adaptability to new categories, overcoming the limitation of standard caption-based approaches
Solution Approach 2:
The patent creates a universal training framework using information matrices that can be applied across multiple domains and use cases. The structured representation serves multiple functions: enabling accurate object detection, facilitating adaptation to new categories, and integrating domain-specific knowledge, thereby replacing the need for domain-specific model variants
4Productivity
If machine learning models are trained with limited training data in novel situations, then the model can be trained quickly, but the model cannot achieve precise predictions for objects never seen before
Solution Approach 1:
The patent changes the parameter representation by transforming limited training data into structured information matrices that encode rich domain-specific relationships and attributes. This parameter transformation enables the model to achieve precise predictions for novel objects even with limited training data, as the structured matrices provide comprehensive information about object properties and relationships
Data Source
AI summary
A method is for training a base model for object detection, trajectory prediction, and/or motion planning of a vehicle. The method includes providing a training data set of image data, with each piece of image data having information about at least one driving scene from a point of view of the vehicle, and providing a knowledge graph including domain-specific knowledge of the at least one driving scene. The method further includes optionally partitioning the image data into a plurality of image sections, and generating information matrices corresponding to the image sections by assigning domain-specific knowledge about the at least one driving scene extracted from the knowledge graph and/or directly from the image data to the plurality of image sections of the image data. The method also includes training the base model based on the information matrices.